Prompt Compression

ReviewAudited by ClawScan on May 10, 2026.

Overview

This looks like a local prompt-compression tool, but it asks users to rely on missing/unreviewed scripts and would broadly rewrite persistent agent prompt files.

Install only if you can inspect the actual scripts first. Use a git-controlled repo, narrow the configured targets, back up important prompt files, and review all diffs before running another agent loop.

Findings (2)

Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.

What this means

Following the install/run instructions could fail or lead the user to rely on unreviewed scripts from outside the provided artifacts.

Why it was flagged

The skill describes executable installer and compression scripts, but the provided package says no code files are present and the manifest does not include these scripts. That makes the core executable behavior unavailable for review.

Skill content
`scripts/trinity-compress.sh` — compression script ... Installers: `scripts/install.ps1` ... `scripts/install.sh`
Recommendation

Do not run installer or compression scripts unless they are actually present in the package and you have reviewed them. The publisher should include the referenced scripts, declare required binaries, and provide a real install spec.

What this means

A compression pass could weaken, distort, or remove important instructions in prompts or skills, and those changes may affect later agent runs until undone.

Why it was flagged

The configuration targets persistent agent and system prompt files, including all skill SKILL.md files under broad globs. Rewriting these files can change future agent behavior across tasks.

Skill content
"targets": [".ai-workflow/skills/**/SKILL.md", "skills/**/SKILL.md", "scripts/ralph/CLAUDE.md", "claude.md", "openai.md", ".cursorrules", "system_prompt.md"]
Recommendation

Run only on a clean, version-controlled repo; narrow the target list; exclude safety/security-critical prompts and third-party skills; review diffs before using the modified prompts; and verify undo works.